steve
/
factor
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

math.integers, bignum/f, improve performance. 

This changes avoids looping many times if the denominator is a power of
2. After this change, the implementation matches the linked sbcl
algorithm.  This was probably a mistake done when porting the algorithm.
Basically, as an optimization, all trailing zeros are removed from the
base2 representation of the denominator to have smaller bignums to
divide. But the previous factor implementation didn't take this into
account when making the initial guess of the shift of the numerator to
obtain a result in the range [2^54-1,2^53]. The loop would then correct
the initial guess by a factor of 2 at each iteration, so it would run as
many iteration as the denominator base2 power reduction, instead of only
a few times. For pathological cases, the speed up is huge (10^4):
1 1000 2^ bignum/f
db4
Jon Harper 2015-07-26 18:07:10 +02:00 committed by John Benediktsson
parent cc6d90a096
commit 5424ad5586
1 changed files with 8 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
core/math/integers/integers.factor
View File

@ -124,10 +124,15 @@ M: bignum (log2) bignum-log2 ; inline
: (epsilon?) ( num shift -- ? ) : (epsilon?) ( num shift -- ? )
dup neg? [ neg 2^ 1 - bitand zero? not ] [ 2drop f ] if ; inline dup neg? [ neg 2^ 1 - bitand zero? not ] [ 2drop f ] if ; inline
: scale-numerator ( num den -- epsilon? num' scale )
over [ log2 ] bi@ - [
54 + [ (epsilon?) ] [ shift ] 2bi
] keep ; inline
: pre-scale ( num den -- epsilon? mantissa den' scale ) : pre-scale ( num den -- epsilon? mantissa den' scale )
2dup [ log2 ] bi@ - scale-denonimator [
[ neg 54 + [ (epsilon?) ] [ shift ] 2bi ] [ scale-numerator ] keep swap
[ [ scale-denonimator ] dip + ] bi-curry bi* ; inline ] dip swap - ; inline
! Second step: loop ! Second step: loop
: (2/-with-epsilon) ( epsilon? num -- epsilon?' num' ) : (2/-with-epsilon) ( epsilon? num -- epsilon?' num' )